| Objective: Based on the requirements of tertiary prevention and precision medicine,in this study,we focused on the X-ray phase contrast microscopic imaging of liver tumors,to explore the application of vascular skeletonization analysis and machine learning method in liver tumor classification and prediction,in order to improve the diagnostic efficiency and accuracy of liver tumors as early as possible.This is of great significance for early detection,early diagnosis and early treatment advocated by tertiary prevention.Methods:Human tumor samples were collected from the Department of hepatobiliary surgery,a 3A hospital affiliated to Xinjiang Medical University.8 SPF grade male C57BL/6 mice were selected and raised in barrier environment for two weeks,the liver and brain tissues were removed after they were killed.All the samples were fixed in 10% neutral formalin,and then stored in a refrigerator at 4℃,the experimental samples were dehydrated with 40% ～95% ethanol solution gradually.Then,the BL13W1 biomedical imaging line station of SSRF was used for X-ray phase contrast microscopy and HE staining sections were made also.We compared the traditional HE staining sections with the X-ray phase contrast microscopic images;We used Image J,Amira,Image Pro Plus and MATLAB software for image processing、three-dimensional reconstruction、microvascular network analysis and vascular skeleton information extraction to obtain three-dimensional quantitative feature information;We used Matlab software to extract the two-dimensional quantitative features of X-ray phase contrast microscope image,such as gray histogram,gray level co-occurrence matrix,gray level gradient co-occurrence matrix,wavelet transform and Tamura texture;Then,we used the area under ROC curve and principal component analysis to screen and reduce the dimension of many features;Finally,R software is used to extract the two-dimensional and three-dimensional quantitative information of liver tumor,and machine learning methods such as decision tree,random forest and support vector machine are used to carry out classification research and classification effect evaluation.Results: It is found that the Xray phase contrast microscope image and its high power image have a high correspondence with HE staining image,and the X-ray phase contrast microscope image can observe the original state information of the tumor from multiple angles.In the research of vessel skeletonization,the extraction method based on endpoint constraint is better than the traditional topology refinement method( <0.05).In the analysis of three-dimensional vascular network,with the growth and development of tumor,the tree structure of normal blood vessels changes、a large number of microvessels proliferate、vascular morphology becomes distorted.The number and diameter of blood vessels have increased(<0.05)、 the volume density of three-dimensional blood vessels have gradually increased(<0.05)and there was also a significant increase in vascular twist(<0.05).In the research of feature classification,we used area under the ROC curve and principal component analysis to screen and reduce the dimension of the feature set.and nine features with Area > 0.75 were selected from 48 two-dimensional features,we mix two-dimensional feature set and threedimensional feature set together for principal component analysis,and get 4 principal components with feature roots >1.00 and cumulative contribution rate ≥ 85%,which form a mixed feature set.Under the two-dimensional feature set,the classification accuracy of decision tree method for four types of liver is 97.5%,91.2%,80.4%,88.9%;the accuracy of random forest method for four types of liver is 98.1%,97.7%,90.3%,94.5%;the accuracy of SVM method for four types of liver is 96.8%,74.2%,60.4%,94.9%;Under the three-dimensional feature collection,the classification accuracy of decision tree method for four types of liver are 97.9%,91.3%,89.4%,93.5%;the accuracy of random forest method for four types of liver are 99.6%,97.8%,96.5%,98.4%;the accuracy of SVM method for four types of liver are 99.6%,62.3%,68.5%,99.5%;Under the mixed feature set,the classification accuracy of decision tree method for four types of liver are 99.9%,95.2%,98.4%,96.9%;the accuracy of random forest method for four types of liver are 99.8%,98.3%,99.4%,99.3%;the accuracy of SVM method for four types of liver are 98.2%,90.4%,83.9%,87.1%.Conclusion: X-ray phase contrast micrograph and its high power image have a high correspondence with HE staining image,and the X-ray phase contrast microscope image can observe the original state information of the tumor from multiple angles.The method based on endpoint constraint also has more advantages than the traditional algorithm in the extraction accuracy,and can provide more accurate three-dimensional feature information,reflecting the occurrence and development of normal liver tissue to liver tumor.The results show that the classification effect of using mixed feature set is better than that of using only two-dimensional feature set and three-dimensional feature set,they have increased by 2%,6%,17%,1% and 1%,10%,8%,1% respectively;The classification effect of decision tree and random forest method is better than that of support vector machine method,they have increased by 1%,17%,18%,0% and 1%,23%,24%,3% respectively.The results of this study not only provide more reference for hepatobiliary surgery and imaging doctors in the auxiliary diagnosis of liver tumor,but also play a positive role in the construction of tertiary prevention and precision medicine. |
PDF Full Text Request